小楼

At last, someone in the robotics industry is starting to measure “how much work got done” instead of “how many units were made.”

On Oct. 29 in Shanghai, AGIBOT introduced a new metric at Innovation Day 2026: Worken (Work + Token), which it calls “effective intelligent work hours” in Chinese.

The formula is straightforward:
Worken = number of robots actively working × hours of autonomous operation × output efficiency factor

1️⃣ Only robots actually carrying out tasks count; inventory and showroom displays do not
2️⃣ Only autonomous working time counts; powered-on standby and remote operation by a person nearby do not
3️⃣ Output is measured against a task benchmark, such as the qualified output of a human worker under the same standards

At the same event, AGIBOT also unveiled the G2 Max industrial robot (21 degrees of freedom, capable of carrying 40 kg with its dual arms), the GO-3 Preview embodied foundation model, and plans for Link-U OS 1.0 to be fully open-sourced and donated to the OpenAtom Foundation before the end of the year. The company said its 20,000th unit rolled off the production line in September.

My take:
This is the same shift we’ve seen in AI, from “parameter races” to “measuring token output.” Sooner or later, Physical AI valuations will move beyond shipment-volume narratives and focus on “how much qualified work each machine actually does in a year.” Whoever is first to make that figure auditable will have pricing power.

But let’s stay grounded: Worken is currently just AGIBOT’s proposal, not an industry standard. How task benchmarks are set and who audits the data have yet to be worked out. Until we see third-party-verifiable Worken data, treat it as a useful framework—not a KPI to take on faith.

Projects building DePIN / robotics data networks in crypto can take note: settling payments by effective work hours is much more credible than issuing tokens based on the number of devices.

Sources: AGIBOT official website / PR Newswire press release dated Oct. 29, The Robot Report
#AI #机器人 #PhysicalAI
For informational purposes only; not investment advice.

For AI agents entering Web3, the first gateway isn’t a wallet—it’s a chat box.

DFINITY announced today that Internet Computer is now an official connector for Claude. Find it in Claude’s connectors, connect once, and you can use natural language to query app data on $ICP and call methods. ChatGPT already had an official plugin; now Claude has one too.

I took a look at Claude’s official marketplace page and DFINITY’s documentation. Here are a few key points:

1️⃣ Identity is handled through Internet Identity. When logging in, you choose the permissions yourself: read-only mode (Questions only) or an interactive mode (Actions & questions). You can also choose how long the session lasts.
2️⃣ Permissions are time-limited and expire automatically. You can revoke them at any time, and deleting the connector disconnects it immediately.
3️⃣ What it can do: discover an app’s canisters, query data, and perform authorized “non-financial” updates.
4️⃣ What it can’t do: make financial transactions, top up canisters, or create or deploy canisters. The official documentation is very clear about this.
5️⃣ It can only access canisters that an app has explicitly declared open to agents. Anything without a published declaration cannot be read or accessed through method calls.

My take: this design is fairly restrained. It separates “AI clicks buttons for you” from “AI moves money for you,” tightens permissions by default, and only grants access when you opt in. That’s much more sensible than immediately promising fully autonomous agent trading.

But don’t get carried away: until it can move money, this is more of a developer/productivity tool. Whether it’s a gimmick depends on how many apps in the ecosystem are actually willing to publish agent declarations.

One more reminder: there have been plenty of recent incidents involving malicious plugins and extensions, so only connect to official sources.

Sources: DFINITY’s official X account (10-30), Claude’s official connector page, internetcomputer.org/icp-mcp

#ICP #AIAgent #MCP

Not investment advice. DYOR.

African fintech OPay rings the bell at the NYSE tonight—on the same day U.S. tech stocks took a collective nosedive 😅

OPay has officially priced its IPO under the ticker OPAY:

1️⃣ Offering 16 million ADSs at $19 each, raising about $304 million
2️⃣ Priced at the bottom of the $19–21 range, implying a market cap of about $4 billion (Renaissance Capital / Reuters)
3️⃣ Underwriters have a 30-day option to purchase an additional 2.4 million ADSs; closing is expected on November 2
4️⃣ Backed by SoftBank; Citi, Deutsche Bank, and Standard Bank are leading the underwriting

So what exactly does the company do?
👉 Headquartered in Singapore, with Nigeria as its main market: it’s Nigeria’s No. 2 app across all categories by daily active users
👉 With a 9.9% share of monthly active users, it’s the largest financial app in Africa and the Middle East
👉 In Indonesia, it matches borrowers with lenders; in Egypt, it provides merchant acquiring; in Pakistan, it offers acquiring and wallets

My take:
Pricing at the bottom of the range is nothing to be ashamed of. Getting priced on schedule and going public as planned on a risk-off day shows there are real buyers willing to bet on the “digital banking in emerging markets” story.

There’s also a useful point of comparison for crypto: Nigeria is already one of the most active markets for stablecoins and P2P use. Traditional fintech and on-chain dollars are competing for the same users. Whoever makes “opening an account + transferring money + accepting payments” part of people’s everyday routine will capture the next billion-user wallet opportunity.

My long-term view on U.S. stocks remains the same: short-term pullbacks are noise. The fact that capital markets are willing to welcome new companies is itself a sign of confidence.

Sources: OPay’s official pricing announcement (PR Newswire, 10-30), Renaissance Capital, Reuters
Not investment advice. IPOs are volatile; DYOR.

Base is going from “one block every 2 seconds” straight to “5 blocks per second.” This isn’t a preconfirmation like Flashblocks—it’s a real, canonical block.

The Denim hard fork date is set: activation on Base mainnet is scheduled for November 5 at 18:00 UTC (2:00 a.m. Beijing time on November 6). It’s already been running on the Sepolia testnet since October 21. Node operators must upgrade to base v1.5.1 before activation.

Here are the key changes to watch for:

1️⃣ Native 200 ms block production: Every block has its own block number, hash, state root, and receipts, and goes through the full unsafe→safe→finalized lifecycle.

2️⃣ Flashblocks is being retired: The WebSocket stream will stop at activation. Apps and indexers that rely on pending state need to migrate to canonical blocks and RPC subscriptions (such as newHeads).

3️⃣ The easiest trap to fall into: block.number will advance 10 times faster in real time. Any contract logic that uses “block counts” to measure durations will have those durations shrink to one-tenth of their original length after activation. Hard-coded block-count constants in immutable contracts can’t be changed after activation. block.timestamp is unaffected and will still advance in whole seconds.

4️⃣ A new BaseTime predeploy provides onchain access to millisecond-level time. RPC will include new timestampMs / blockTimestampMs fields; the existing timestamp field remains unchanged for backward compatibility.

5️⃣ B20 transfer rules are getting stricter: Transfers, mints, and seizes to the token’s own address will revert. PolicyRegistry will support NOT-inverted policies.

My take: The era of L2s competing on TPS is long past. The focus now is on the “feel” of confirmations and the execution speed of DeFi and trading apps. By making preconfirmation native to blocks, Base is turning the sub-second experience from an “add-on” into a “default property of the chain.”

But the real risk isn’t performance—it’s migration. If older contracts that use block counts to calculate lockup periods, reward cycles, or auction durations aren’t addressed before November 5, it could cause a silent parameter incident. If you have positions on Base, check whether the protocols you use have announced a migration.

The L2 speed race in the $ETH ecosystem is entering its next round.

Sources: Base official documentation, docs.base.org/upgrades/denim; GitHub base/base v1.5.1 release notes; The Defiant.

This is not investment advice.

Curve’s first move beyond EVM wasn’t to Solana, but to Stellar 🤔

$CRV announced officially on 10-29: StableSwap is now live on Stellar mainnet. This is Curve’s first native non-EVM deployment.

A few key facts:
1️⃣ Stellar doesn’t run EVM, so Vyper contracts couldn’t be ported directly. The team rewrote StableSwap from scratch on Soroban (Stellar’s smart contract platform), with support from SDF (Stellar Development Foundation). The pricing math hasn’t changed.
2️⃣ SDF proactively invited Curve to deploy there. In essence, Egorov said there is demand, and it’s coming from institutions.
3️⃣ For now, it’s a lite version: stellar.curve.finance is still in beta, with the first pool now open for swaps and adding LPs. Pool creation is permissionless; the frontend pool-creation feature is coming next.
4️⃣ Fees work just like on other chains: swaps on Stellar also earn a share for Curve DAO / veCRV.
5️⃣ Next up is FXSwap, designed for relatively volatile pairs such as “USD stablecoin ↔ EUR stablecoin.” Development is mostly complete and it’s awaiting an audit; no timeline has been announced.

My take:
The significance of this move isn’t today’s TVL, but two things—
First, Stellar has a bunch of compliant stablecoins and tokenized funds, but what’s missing is a sufficiently deep venue for trading and settlement. Lending protocols will only dare to accept these assets as collateral once Curve pools are available.
Second, once FXSwap goes live, the $XLM ecosystem will have proper infrastructure for on-chain foreign exchange. That’s what Stellar was made for.

Of course, with a lite version, beta, and just one pool, whether liquidity takes off depends on whether issuers and institutions actually put real money in. Just filtering, not educating—making a note for now and waiting for the FXSwap audit and pool data.

Sources: Curve official blog, news.curve.finance; Stellar official press release; The Block (10-29)
Smart contract and depeg risks remain. This is not investment advice.

Kioxia’s earnings report falls into the “profits exploded, but guidance gives you pause” category.

Kioxia (TSE: 285A) announced FY2026 Q2 results (July–September) late on 10-30:
1️⃣ Revenue of ¥2.48 trillion, up 40% quarter over quarter and about 5.5 times the same period last year
2️⃣ IFRS operating profit of ¥1.995 trillion, with an operating margin of 80.4% (vs. 19.2% in the same period last year)
3️⃣ Gross margin of 82.8%; net profit attributable to owners of the parent of ¥1.356 trillion
4️⃣ Both revenue and operating profit came in above the company’s own guidance from July
5️⃣ First-ever dividend: an interim dividend of ¥90 per share (on a pre-split basis), payable from 12-8; according to Bloomberg, this is the first in the company’s history

The drivers are clear: generative AI data centers need SSDs, average NAND prices have surged, and bit shipments are also rising.

The issue is next quarter 👇
The company guided for October–December revenue of ¥2.5 trillion and operating profit of about ¥2 trillion, essentially flat from Q2, citing a stronger yen and restrained pricing for data center customers. The market expectation cited by Newsquawk was operating profit of ¥2.35 trillion; Nikkei reported net profit guidance of ¥1.36 trillion, below QUICK’s consensus estimate of ¥1.59 trillion.

My take:
In this memory cycle, SK hynix has an 80% operating margin, SanDisk an 85% gross margin, and Kioxia an 80.4% operating margin. AI-driven demand for memory is translating into real money. But the market is trading on “acceleration quarter over quarter,” and when guidance goes flat, the share price often takes a hit first.

Flat ≠ bad. ¥2 trillion in operating profit in a single quarter is still extraordinary; it’s just a shift in the script from “surging every quarter” to “holding steady at a high level.”

Two things to watch next: whether NAND prices have room to rise further, and the yen exchange rate. One reminder: an 80% margin itself is a sign that this is still a cyclical business.

Sources: Kioxia’s 10-30 earnings release and dividend announcement (TDnet), Nikkei, Newsquawk, Quartz. Not investment advice.

Gemini’s Winklevoss brothers have been clashing head-on with the Zcash community on X over the past couple of days 😂

The backstory: On 10-27, Zcash co-founder Sean Bowe published a draft proposing to end the automatic 8% of each block going to Zcash Community Grants (ZCG) and improve the Network Sustainability Mechanism. By his estimate, ZCG still has about $125M left after accounting for grants already committed—enough to last for years (using $1,300/ZEC as an illustrative price). (CoinDesk)

1️⃣ The current setup: 20% of block rewards goes to ecosystem development; 8% goes to ZCG (decisions made by a five-member committee, with no separate vote by holders); and 12% goes to retroactive grants voted on by holders.

2️⃣ Cameron’s remarks on 10-30, paraphrased: Shareholder voting has been used in capital markets for hundreds of years, and coinholder voting is its on-chain equivalent. Zcash’s mission is privacy, not experimenting with governance. It’s time to end ZCG.

3️⃣ Tyler piled on: Comparing ZCG to mining doesn’t hold up. Miners compete in the market with real money on the line; a committee handing out money doesn’t.

4️⃣ The opposing view: Grants are a reasonable way to support positive externalities for the ecosystem; some also worry that holder voting naturally favors large holders.

My take: At its core, this is about who spends treasury funds and who is accountable for the results. With a pool worth around $125M, it’s getting harder for a committee-based system to demonstrate its efficiency. But holder voting means whales get a say—and the Winklevoss brothers are themselves large holders, so their own interests may shape their views.

The timeline matters more than the rhetoric: The draft can only be implemented with a future network upgrade, and no changes have been approved so far. Holders plan to collect proposals for the next round from 10-30 to 11-13, with voting from 12-17 to 12-29.

Just filtering, not financial education: $ZEC holders can keep an eye on the forum and upgrade proposals. Not investment advice.

What did Aave do after the Binance hot wallet incident? Not just talk—it froze the markets.

According to LlamaRisk’s official update on the Aave governance forum (10-30): Starting at 20:24 UTC on 10-29, after the attacker transferred a large amount of assets out of the Binance hot wallet, an associated address deposited $60 million in $USDT into Aave V3 Ethereum Core within about 40 minutes and borrowed 19,762 WETH (worth about $44.6 million at the time). The position was overcollateralized and posed no bad debt risk to suppliers. The address did not interact with Aave again afterward.

Aave Labs, TokenLogic, Certora, and LlamaRisk opened a war room. The guardian multisig carried out freezes between 21:42 and 23:04 UTC:
1️⃣ Aave V3: 175 reserves across 16 chains
2️⃣ Aave V4: 117 reserves across 4 networks
After the freeze: new deposits and borrowing were paused; withdrawals, repayments, and liquidations continued as usual. The Horizon RWA market was not frozen—the collateral is permissioned, so the attacker could not use it.

Today (10-30), LlamaRisk also proposed a Risk Stewards plan: before unfreezing, reduce large borrow/draw caps to “the amount borrowed before the incident + about 2%.” Roughly, this would cut unused borrowing capacity by about $3.7 billion on V3 and about $130 million on V4. For example: the V3 Core WETH cap would drop from 2.4 million to 1.79 million; $USDT from $3.09 billion to $2.56 billion. The reasoning is straightforward—if hundreds of millions in unused capacity remain when the markets are unfrozen, the attacker could immediately use those assets to build another position.

Analyst’s takeaway: When a CEX is compromised, the biggest lending pools can instantly become a “cash-out” channel. A freeze is a reversible guardian action, faster than going through five days of governance to cut caps. Next, watch whether proposals for staged unfreezing and permanent speed limits such as velocity limits are implemented. $AAVE $ETH

Source: Aave governance forum, LlamaRisk, “Binance Hot Wallet Exploit: Aave’s Response” and “Risk Stewards borrow/draw cap reductions 2026.10.30.” Not investment advice.

PhotonPay just closed a $40 million Series C.

The round was led by Blockchain Capital, with new investor SBI Group and existing investors IDG, GL Ventures, and Lightspeed Faction participating. The original announcement is on PR Newswire.

What does it do? A financial OS built around stablecoins—multi-asset wallets, cross-border settlement, card issuance, and local clearing, connecting on-chain liquidity to businesses’ everyday spending and payments.

The hard numbers (as of August 2026): $65 billion in annualized transaction volume; 34 payment/financial licenses; over 200,000 business customers across 200+ countries and regions; 4.4 million cards issued in the past 12 months, up 275% year over year.

CEO Lewison Chen put it in a nutshell: stablecoins solve how money moves; AI solves how money gets put to use—and commerce is the endgame. In May, the company also did a live-network demo of agentic payments with Mastercard.

In the stablecoin infrastructure space, money is still flowing to players who can connect to fiat clearing networks. This is for screening, not education, and does not constitute investment advice.

TotalEnergies has also reported its third-quarter results, rounding out this European oil majors’ refining boom.

Adjusted net income was $9.1B (Q2: $6.0B, up 51%; roughly double year over year), CFFO was $12.3B, and adjusted EPS was $4.03. Gearing fell sharply to 9.1% (from 13.1% in Q2). In line with its target of returning 40% of cash flow, the board approved $2.5B in buybacks for each of Q4 and 2027Q1, while maintaining the third-quarter interim dividend at €0.90.

A few key points:
1️⃣ Record downstream performance—adjusted net operating income of $4.1B, refinery throughput up nearly 10% quarter over quarter, benefiting from historically high refining margins in Europe and globally (European Refining Margin Marker: $33.2/b)
2️⃣ Integrated LNG at $1.9B (more than doubling quarter over quarter); average LNG prices up 20%, with trading outperforming expectations
3️⃣ E&P production of 2.474 Mboe/d, impacted by Middle East conflicts by around 5%; excluding the impact of the Middle East, production was up >4% year over year
4️⃣ Also recorded around $1.9B in impairments (including offshore wind and the Hanwha petrochemical joint venture), while reported net income still came in at $8.3B

Same playbook as Shell this morning and BP / Exxon / Chevron this evening: refining margins and gas prices boosted cash flow, with debt repayment before buybacks. The U.S. energy sector value chain is still making money. This is not investment advice.

The Canadian government is taking a direct stake in an AI company.

On 10-29, Prime Minister Mark Carney announced that the federal government will invest up to CAD 1.6 billion in Toronto-based Cohere in stages, on commercial terms. The announcement came alongside a Series E led by Schwarz Group, targeting up to CAD 7 billion at a post-money valuation of CAD 35.7 billion. New investors include Bell Ventures and Bullhound; existing investors NVIDIA, Inovia, and Prosperity7 are also participating.

The largest private financing in Canadian history. The conditions are set in stone:
1️⃣ The company and its headquarters must remain in Canada
2️⃣ Invest at least CAD 1.7 billion in local computing capacity by 2030
3️⃣ Create at least 1,000 high-skilled technical jobs in Canada

The bigger picture: Canada wants a third path to frontier AI beyond the US and China. Cohere just merged with Germany’s Aleph Alpha, focusing on controllable data for enterprises and governments, including in regulated sectors. I wrote about Mistral’s open-source release this morning; tonight, Canada is putting money in and taking a stake—the sovereign AI story is becoming clearer. NVIDIA sells the shovels, and doesn’t miss out on either side.

Filtering, not educating. Not investment advice.
Source: pm.gc.ca + Cohere PR Newswire

SK hynix Q3: Revenue tops KRW 100 trillion for the first time.

This AI memory boom isn’t just a PowerPoint story—it’s showing up on the income statement.

Official figures (10-28 IR):
1️⃣ Revenue: KRW 104.6 trillion (about $75.5B), up 328% year over year and 32% quarter over quarter
2️⃣ Operating profit: KRW 83.3 trillion; operating margin: 80% (Q2: 76%)
3️⃣ Net profit: KRW 39.8 trillion

HBM4 volume is expected to ramp up in Q4, surpassing HBM3E in both shipments and revenue to become the mainstream product. The company has signed long-term agreements with 14 major customers, covering more than half of its capacity for 2027–2028. The company expects the memory shortage to potentially last until 2030.

Shareholder returns are substantial too: about KRW 100 trillion in total this year (about $75B)—at least KRW 30 trillion in cash dividends and at least KRW 60 trillion in share buybacks and cancellations (including KRW 40 trillion completed in August). The Q3 dividend is KRW 20,000 per share, half as a performance dividend and half as a special dividend from gains on its Kioxia investment.

I wrote this morning about AI driving memory demand at SanDisk; here, the HBM picks-and-shovels maker is making the numbers real. The people selling shovels have steadier cash flow than the miners.

Screening only, not educating. This is not investment advice.
Source: SK hynix Newsroom 3Q26

The last weak spot in AI video just got a little stronger—ElevenLabs takes another step forward 🎧

On 10/29, ElevenLabs launched SFX v3, its next-generation text-to-sound-effects model, available in both ElevenCreative and the API.

A few of the upgrades highlighted by the company:
1️⃣ It can understand prompts that describe several sounds at once. For example, “crowd cheering at a stadium, a ball bouncing, spectators booing.” Previously, you had to generate each sound separately and combine them; now, one prompt produces the whole sequence.
2️⃣ Sound effects with music and vocals sound more convincing (the company’s example: “Gospel choir singing hello”).
3️⃣ Fewer clipping and unwanted noises in high-frequency sounds (gongs, wind chimes, whistles).
4️⃣ Seamless looping is supported, with clips up to 30 seconds long. In the API, just change the model_id to eleven_text_to_sound_v3; v2 is still the default.

A few reasons to keep things in perspective:
· The official docs say access is limited, and some regions and accounts may get an “Unknown model” error.
· The claim that it’s “better than v2” comes from the company’s own benchmark, and the data hasn’t been made public.
· The docs also recommend comparing v2 and v3 on the same prompt before switching. Even precise requests like “strike a triangle 4 times” aren’t guaranteed.

My take: AI video has been in a race to improve visuals these past two years, but sound is what has always taken the most human effort—music, ambience, and foley. As models take over sound effects, the cost of making short films and game demos for small teams takes another step down.

The shovel sellers keep making money, and the shovel users keep chasing efficiency. Don’t just filter—learn. Try the prompts yourself; your ears are more honest than a benchmark.

Source: ElevenLabs official blog and developer documentation (2026-10-29)
#AI #ElevenLabs #AIGC

America’s two oil giants reported on the same day, and came to the same conclusion: refineries are printing money 🛢️

10-30 ExxonMobil and Chevron reported Q3 results together. BP had its moment yesterday; today it was the turn of the homegrown American players.

1️⃣ ExxonMobil: Adjusted EPS of $4.23, versus LSEG expectations of $4.08; adjusted earnings of $17.5B. Operating cash flow was $25.7B, and free cash flow was $19.1B. The company returned $9.4B through dividends and buybacks this quarter, and also announced it would continue buying back $20B in shares in 2027.

2️⃣ Chevron: Adjusted EPS of $5.28, versus expectations of $4.89; GAAP earnings of $9.9B, compared with just $3.5B in the same period last year. U.S. refineries processed a record 1.09M barrels of crude per day, with utilization near 98%. The quarterly dividend was $1.78, and net debt fell by $7.8B in a single quarter.

3️⃣ The common thread: upstream is actually weakening. Exxon’s adjusted upstream earnings fell 20% quarter over quarter, while Chevron’s upstream earnings were down about 30% (Reuters), as oil prices were lower than last quarter. Downstream carried the quarter: Exxon’s adjusted downstream earnings came in at $9B, up 120% quarter over quarter. Conflict in the Middle East, along with attacks on Russian and Ukrainian refineries, has tightened global diesel supplies and sent refining margins soaring.

My take: This is money the cycle is handing out, not money from oil prices. The real question is what each company does with it. BP chose to pay down debt first; Exxon is sticking with buybacks while keeping its 44-year streak of dividend increases intact; Chevron is doing both—paying down debt and buying back shares.

One dry joke: AI is hungry for electricity, diesel is in short supply, and in the end, a bunch of old-school refineries are making a killing.

That’s where the resilience of the U.S. stock market machine comes from: tech is burning through capex, while energy is printing FCF. Both sides of the story are about cash flow.

Sources: Official ExxonMobil and Chevron earnings releases dated 10-30 (BusinessWire); consensus estimates from Reuters/LSEG.
Not investment advice.